Product1 distinct publisher3 min readUpdated
RoboScience's REX G1 trades demo-reel language for numbers a warehouse engineer can argue with. What it still does not publish is price, throughput, or any outside validation.
The Product Desk · Product desk

Compiled by The Product DeskSomething wrong?How this is made
RoboScience, a Beijing embodied intelligence company, released its first robot on August 17: the REX G1, a wheeled humanoid pitched at logistics, manufacturing, retail and eventually domestic work [1][2]. The notable thing is not the shape but the disclosure, because the company has published the kind of figures a facilities engineer can check against a doorway and a shift plan rather than a video.
Start with the mechanics. The REX G1 carries 22 degrees of freedom, allocated as two in the head and neck, seven per arm, two in the waist, one lifting axis and three in the wheeled chassis [3]. That accounting closes exactly, which is a low bar and one that marketing copy in this category often fails to clear [4]. The two arms together are rated for 10 kg, or 22 pounds [5], which averages 5 kg per arm if a load is split evenly [6]. That is tote-lid territory, not full-tote territory, and it sets the boundary of the job more clearly than any claim about general capability.
The dimensions are written as if someone measured an existing aisle. The robot is 540 mm wide, adjustable from 540 to 640 mm in length, and ranges from 1,100 to 1,700 mm in height, with a working envelope between 0.1 and 2 meters [7]. That height range implies 600 mm of vertical travel from the lifting axis [8], and 540 mm is about 21.3 inches of width to fit through [9]. RoboScience frames this as the opposite of the usual retrofit demand for wider aisles, more turning space or altered shelving [10]. Perception is 360-degree stereo vision with adaptive lighting, intended for deep cabinets, reflective shelving, packaging film and low light without added fixtures [11]. The vision system can also read storage-location codes after a grasp to confirm placement [12].
Two numbers do the most work for a buyer. The first is repeatability of plus or minus 0.1 mm [13], reported without a stated test payload, reach or measurement standard, which is where such figures usually move [14]. The second is the battery: up to four hours per charge, hot-swappable without shutting down, with work resuming in as little as five seconds [15]. Taken at face value, five seconds against a four-hour run is roughly 0.035 percent of the cycle spent swapping, or about 99.97 percent availability before anything else goes wrong [16]. That is the claim that turns a robot into a shift asset instead of a day-shift experiment, and it also quietly assumes spare packs, a charging bay and someone to do the swap.
The software is called Visics, which RoboScience describes as a general embodied model on a Vision-Language-Object-Action architecture [17]. The company says a world model simulates three-dimensional object trajectories and a separate operation model converts those into control signals [18]. Onboard compute is quoted at 2,070 TFLOPS, with perception, decision and control run at the edge to reduce cloud dependence [19]. RoboScience also claims "walking, lifting, and grasping at the same time," with zero-space control holding end effectors steady in tight spaces so the machine does not stop and reposition between actions [20].
What to watch: nothing here is priced, dated for shipping, or attached to a named customer, and every figure is the vendor's own [21]. The missing specs are the operational ones - picks per hour, mean time between failures, payload-derated accuracy, and how many spare batteries a two-shift line needs. The category has moved from adjectives to numbers. The next move is from vendor numbers to somebody else's.
Follow any of these and your For You feed starts watching them — no settings page required.
Ranked by verification strength, evidence, and original report placement.
RoboScience, a Beijing-based embodied intelligence company, introduced its first robot, the REX G1, for use across logistics, manufacturing, retail and domestic environments.
The REX G1 was released on August 17; it is a wheeled humanoid that combines movement and manipulation with RoboScience's Visics model.
The REX G1 has 22 degrees of freedom: two in its head and neck, seven in each arm, two in its waist, one lifting axis and three in its wheeled chassis.
The REX G1's two arms can carry a combined load of 22 pounds (10 kg).
The REX G1 measures 540 mm wide, with length adjustable from 540 to 640 mm, height ranging from 1,100 to 1,700 mm, and a working range between 0.1 and 2 meters.
Many warehouse robots require changes such as wider aisles, additional turning space, or adjusted shelving; RoboScience designed the REX G1 to fit into existing facilities instead.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Single vendor-sourced spec disclosure
One trade article carries the entire cluster, and every figure in it is attributed to RoboScience. The published numbers are checkable only for internal consistency — the DOF breakdown sums to 22, the height range implies 600 mm of travel, the swap-to-runtime ratio is arithmetically sound — which is better than pure demo-reel language but far short of independent measurement. The headline performance figures (±0.1 mm repeatability, 2,070 TFLOPS, simultaneous locomotion and manipulation) arrive without test conditions, and nothing has been validated by a third party.
Launch announcement only
Adoption evidence stops at the product release itself. There is no named customer, no pilot deployment, no shipped-unit count, no price and no availability date in the supplied material, and the application list (warehouse, manufacturing, retail, future home services) is stated as intent rather than as delivered work.
Numbers published, performance still unproven
The gap is positive but modest. RoboScience has done more than most humanoid announcements by publishing an argued spec sheet, and the arithmetic holds where it can be checked. But the claims that would actually justify workplace deployment — ±0.1 mm repeatability, brownfield fit without facility changes, simultaneous walking-lifting-grasping, a 'next-generation productivity partner' framing across four industries — are asserted with no throughput data, no site trial and no outside validation, while adoption is a single release event.
Vendor launch narrative, unchallenged
Every claim originates with the company selling the product, at the moment of its first product launch, and the article reproduces the vendor's positioning language directly without a counter-source, competing spec comparison or independent test. That is a strong promotional incentive structure. It is partly mitigated by the fact that the vendor chose to publish falsifiable numbers rather than only qualitative capability claims.
Low — one publisher, one vendor voice
Confidence is limited by structure rather than by contradiction: nothing in the cluster is contested, but there is only one publisher, one underlying vendor source, and no measurement or deployment evidence to triangulate against. The specification facts can be reported with reasonable confidence as published claims; the performance and market-fit assertions cannot.
product
Foundation says it demoed a humanoid to DHS. DHS says there is no contract.1 distinct publisher
product
A $1,688 home humanoid ships this fall. Read the payload spec, not the price tag.1 distinct publisher
product
The robot is off the shelf. The audience is in the voice layer.1 distinct publisher
product
The installed robot base, not the humanoid, will remodel the factory first1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.